My favorite is Multiple Regression in Behavioral Research: Explanation
and Prediction (3rd ed.) by E.J. Pedhazur (1997), Harcourt Brace.
Multicolinearity causes no serious problem in prediction, as long as
you are not interested in comparing the predictor variables for relative
importance (i.e., ordering the variables for "importance"). Therefore,
maximizing R-sq is a fine goal, but do not attempt to say which variable is
"most important, second most important, etc." in any absolute sense. If I
recall correctly, the estimators are also unbiased under conditions of
multicolinearity.
Burke Johnson
University of South Alabama
>>> "Van den Poel, Dirk" <[log in to unmask]>
03/25/98 05:05am >>>
Dear list-members,
I am looking for good references on the following topic: impact of
multicollinearity on the predictive performance of statistical (and
non-statistical: e.g. machine learning) techniques
I am a PhD student in marketing and I am especially interested on the
impact on predictive performance, not so much on hypothesis testing.
Any related comment is greatly appreciated,
Dirk
Dirk Van den Poel
K.U.Leuven
Department of Applied Economics
Naamsestraat 69
B-3000 Leuven
Belgium
Phone: + 32 16 32 69 45
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|